GPT-4.5 Achieves 73% Success Rate in Turing Test with Persona at UC San Diego

In a recent study at the University of California, San Diego, OpenAI's GPT-4.5 has demonstrated a remarkable ability to mimic human communication, achieving a 73% success rate in a three-party Turing Test when adopting a specific persona. This suggests that in many cases, participants were more likely to believe that GPT-4.5 was human compared to the actual human participant. The experiment, which involved nearly 300 participants, evaluated GPT-4.5's capacity to engage in text-based conversations and convince interrogators of its human-like nature. When prompted to adopt a persona, GPT-4.5 significantly outperformed its baseline performance, where it only convinced 36% of participants without specific instructions. In comparison, OpenAI's GPT-4o achieved a 21% success rate without a persona. The Turing Test, conceptualized by Alan Turing in 1950, assesses a machine's ability to exhibit intelligent behavior equivalent to that of a human. The recent findings highlight the advancements in language models and the impact of prompt engineering on AI's ability to convincingly imitate human interaction. The study also evaluated Meta's Llama 3.1-405B model, which attained a win rate of around 56% with a persona prompt. These results indicate that AI systems are becoming increasingly adept at mimicking human-like conversation, raising important questions about the nature of intelligence and the potential social and economic impacts of these technologies.